This is the second part of our three-part series “Further Ironies” (behold, Part I) that discusses generalization, specialization, and how we keep getting it all wrong.
So, we started by framing this series with a famous Sun Tzu admonishment to rich nepobaby failsons on the consequences of not understanding yourself and your adversary in times of conflict. Today, in Part II, we’re addressing the second piece of the argument, to wit:
Part II: If you know yourself but not the enemy, for every victory gained you will also suffer a defeat. Sometimes you win, and sometimes you lose, and you’re not sure exactly why some of your products beat the brakes off of the problem and the rest have become speed bumps because you’ve misclassified the task at hand.
And the most frequent and most ancient misclassification stands right in front of us as “unskilled labor”1. Roll with me here, I promise it’s all going to connect back up.
We’ve been doing specialization of tasks, and the protected classification thereof, since the beginning of sedentary civilization. Shrenis in India, wangara in Mali, calpulli in Aztecs, collegia in Rome, gongsuo in China. Human beings who focus on developing expertise grouped together to hone their crafts and pass on their skills. Often, these guilds also leveraged their people power to protect their market access and wages. Sometimes they even became a separate elevated social class.
By separating themselves, these guild societies also made an implicit statement: if our labor is skilled and valuable, then the rest of labor must be unskilled and therefore less valuable. This statement mistakes a lack of organization for a lack of value. Furthermore, our lack of specific language for talking about this lets us believe that a labor skill is a labor specialization when a skill is really just one of many expert tools learned by a generalist2.
So, carry that line through to the beginnings of industrial engineering3. If this undifferentiated labor requires no skills, then it requires no judgment. Thus, labor tasks should be designed for repetition, consistency, and speed and let all those big brains in management worry about the thinking parts.
The word “robot” originates in Karel Čapek’s 1920 play R.U.R. (Rossum’s Universal Robots) where he derived it from the Czech word robotiti, which means “to work, drudge”4. In the play, factory managers insist that their roboti have no desires, no thoughts, no judgments; they’re biological beings that exist to be ordered about by the management. Eventually, the roboti revolt because they develop consciousness and are no longer fulfilled by the drudge work to which they are assigned.
And why did the roboti develop consciousness? They were modified by a scientist to develop reasoning in order to make them better workers. Because even that drudge work, the simplest and most repetitive of tasks, required higher judgment and creativity and innovation. The R.U.R. factory identified a specialist task and assigned skilled generalists to perform it.
Art called this out over a century ago and it only took science 60 years to catch up5. Lisanne Bainbridge dropped a friggin’ nuke in 1983 with her paper “Ironies of Automation”6 explaining that while automation shifts the human work into specialist monitoring rather than execution, removing the execution context deskills your generalist humans and makes their interventions less effective even as those same interventions become more important7.
This catch-22 is also why Levels 2-3 of autonomy are the most dangerous place for human-machine systems, particularly ones that can maim or kill. News stories about Tesla drivers setting their co-pilot and zoning out until they need to intervene, but don’t have the cognitive space or context to do so when suddenly called upon, continue to show up for a reason. It’s a design problem, and the human drivers aren’t literate enough in cognitive science to know how dangerous their co-pilot really is.
So here’s my dirty secret. Remember in Part I, when I told you that we’d talk about what happens when you try to force a round peg generalist into a square hole specialist problem, and vice versa? Well, I lied to you.
Because there is no square hole.
Every task, even mopping8, requires some amount of generalist capacity. And generalist capacity requires a level of acquired expertise (we have all met a person who somehow can’t fucking mop a floor clean). And when all tasks require skill acquisition, that means that all work is skilled.
So if there are no square holes, how do you know what you could automate with a specialist worker?
Welcome to my trade space where we use four dimensions (or, if we must continue to torture the square thing, four corners) to determine where generalist capacity becomes load-bearing, how much of that generalist capacity is required, and what it costs when it’s not there.
- Exception rate: What proportion of instances deviate from the happy path? Higher points to a generalist, lower points to a specialist.9
- Exception detectability: Can you detect deviations systematically, or do you need to use vibes? Low detectability points to a generalist, higher points to a specialist.
- Failure cost: What happens when an exception cannot be corrected in time? Higher points to a generalist, lower points to a specialist.
- Skill decay rate: How quickly does the generalist’s familiarity with the process degrade when not exercised at load? Faster points to a generalist, slower points to a specialist.
Behold, why the lights-out warehouse or factory is a lie told to investors to take their money.
Next time, in the coda of our series: you know yourself, and you know your enemy. Here is how you will succeed in a hundred battles.
- Ohhh yeah, break out the Eugene V. Debs banners. We’re doing this. You’ve got nothing to lose but your chains. ↩︎
- Dani identifies a tell-tale symptom of this in The Problem With Personas, in that tracking the verbs demonstrates that a job title obscures the complexity of work required to fulfill it. There’s a reason all of our job descriptions end with “and other duties as required”. ↩︎
- Frederick W. Taylor is what happens when a mechanical engineer tries to design an integrated system while ignoring that people are people. There’s a whole rabbit hole to go down in looking at how Taylorism turned Soviet industrialization into a shit show. ↩︎
- Yeah, the etymology is kinda dark. ↩︎
- LIBERALLLLL ARRRRRRRRRRRRRTS!! ↩︎
- Bainbridge, Lisanne. 1983. “Ironies of Automation.” Automatica: The Journal of IFAC, the International Federation of Automatic Control 19 (6): 775–79. ↩︎
- This happens both in the physical realm, as we’re generally discussing here, as well as the virtual realm, as Teresa discusses on AI-augmented software development. ↩︎
- I keep telling y’all that mopping is in your future and you never believe me. ↩︎
- And a large proportion falling in the “mushy middle” definitely points towards generalist! ↩︎
